void test_vrev64Qs32 (void)
{
int32x4_t out_int32x4_t;
int32x4_t arg0_int32x4_t;
out_int32x4_t = vrev64q_s32 (arg0_int32x4_t);
}
void test(JNIEnv * env, jobject jRoot, jobject jObj) {
int *testSet1 = (int*)malloc(sizeof(int)*DATA_SIZE);
for(uint32_t i = 0; i<DATA_SIZE; i++) {
testSet1[i] = i;
}
clock_t begin = clock();
for (uint32_t i=0; i<DATA_SIZE/4/2; i++) {
int32_t *src = testSet1+i*4;
int32_t *dest = testSet1+DATA_SIZE - 4*(i+1);
int32x4_t tmp = vld1q_dup_s32(src);
int32x4_t destData = vld1q_dup_s32(dest);
int32x4_t rDestData = vrev64q_s32(destData);
vst1q_s32(src, rDestData);
vst1q_s32(dest, tmp);
}
clock_t end = clock();
for (uint32_t i = 0; i<DATA_SIZE/2; i++) {
int t = testSet1[i];
int d = testSet1[DATA_SIZE-1-i];
testSet1[i] = d;
testSet1[DATA_SIZE-1-i] = t;
}
clock_t end2 = clock();
clock_t cost1 = end-begin;
clock_t cost2 = end2-end;
__android_log_print(ANDROID_LOG_DEBUG, "NEON", "last number is %d, acc=%.1fx", testSet1[DATA_SIZE-1], 1.f*cost2/cost1);
free(testSet1);
jclass clasz = env->FindClass("com/tencent/helloneon/BenchListener");
jmethodID method = env->GetMethodID(clasz, "onResult", "(Ljava/lang/String;)V");
std::stringstream out;
out << "benchResult:" << 1.f*cost2/cost1;
env->CallVoidMethod(jObj, method, env->NewStringUTF(out.str().c_str()));
}
inline int32x4_t vrev64q(const int32x4_t & v) { return vrev64q_s32(v); }
void silk_warped_autocorrelation_FIX_neon(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
)
{
if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
} else {
opus_int n, i, lsh;
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
opus_int64 corr_QC_orderT;
int64x2_t lsh_s64x2;
const opus_int orderT = ( order + 3 ) & ~3;
opus_int64 *corr_QCT;
opus_int32 *input_QS;
VARDECL( opus_int32, input_QST );
VARDECL( opus_int32, state );
SAVE_STACK;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
input_QS = input_QST;
/* input_QS has zero paddings in the beginning and end. */
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
/* Loop over samples */
for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
const int16x8_t t0_s16x4 = vld1q_s16( input + n );
vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
}
for( ; n < length; n++, input_QS++ ) {
input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
}
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
/* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */
/* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */
/* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
/* Keep the C code here to help understand the intrinsics optimization. */
/*
{
opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int32 *state_QST[ 3 ];
state_QST[ 0 ] = state_QS[ 0 ];
state_QST[ 1 ] = state_QS[ 1 ];
for( n = 0; n < length + order; n++, input_QS++ ) {
state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
for( i = 0; i < orderT; i++ ) {
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
}
state_QST[ 2 ] = state_QST[ 0 ];
state_QST[ 0 ] = state_QST[ 1 ];
state_QST[ 1 ] = state_QST[ 2 ];
}
}
*/
{
const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
const opus_int32 *in = input_QS + orderT;
opus_int o = orderT;
int32x4_t state_QS_s32x4[ 3 ][ 2 ];
ALLOC( state, length + orderT, opus_int32 );
state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
/* Calculate 8 taps of all inputs in each loop. */
do {
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
n = 0;
do {
calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
} while( ++n < ( length + order ) );
in = state;
o -= 8;
} while( o > 4 );
if( o ) {
/* Calculate the last 4 taps of all inputs. */
opus_int32 *stateT = state;
silk_assert( o == 4 );
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
n = length + order;
do {
calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
input_QS++;
stateT++;
} while( --n );
}
}
{
const opus_int16 *inputT = input;
int32x4_t t_s32x4;
int64x1_t t_s64x1;
int64x2_t t_s64x2 = vdupq_n_s64( 0 );
for( n = 0; n <= length - 8; n += 8 ) {
int16x8_t input_s16x8 = vld1q_s16( inputT );
t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
inputT += 8;
}
t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
for( ; n < length; n++ ) {
corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
}
corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
corr_QC[ orderT ] = corr_QC_orderT;
}
corr_QCT = corr_QC + orderT - order;
lsh = silk_CLZ64( corr_QC_orderT ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
lsh_s64x2 = vdupq_n_s64( lsh );
for( i = 0; i <= order - 3; i += 4 ) {
int32x4_t corr_s32x4;
int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
corr_s32x4 = vrev64q_s32( corr_s32x4 );
vst1q_s32( corr + order - i - 3, corr_s32x4 );
}
if( lsh >= 0 ) {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
}
} else {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
}
}
silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
RESTORE_STACK;
}
#ifdef OPUS_CHECK_ASM
{
opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int scale_c;
silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
silk_assert( scale_c == *scale );
}
#endif
}
static inline int32_t TransformAndFindMaxNeon(int16_t* inre,
int16_t* inim,
int32_t* outre,
int32_t* outim) {
int k;
int16_t* inre1 = inre;
int16_t* inre2 = &inre[FRAMESAMPLES/2 - 4];
int16_t* inim1 = inim;
int16_t* inim2 = &inim[FRAMESAMPLES/2 - 4];
int32_t* outre1 = outre;
int32_t* outre2 = &outre[FRAMESAMPLES/2 - 4];
int32_t* outim1 = outim;
int32_t* outim2 = &outim[FRAMESAMPLES/2 - 4];
const int16_t* kSinTab1 = &WebRtcIsacfix_kSinTab2[0];
const int16_t* kSinTab2 = &WebRtcIsacfix_kSinTab2[FRAMESAMPLES/4 - 4];
uint32x4_t max_r = vdupq_n_u32(0);
uint32x4_t max_i = vdupq_n_u32(0);
// Use ">> 5", instead of "<< 9" and then ">> 14" as in the C code.
for (k = 0; k < FRAMESAMPLES/4; k += 4) {
int16x4_t tmpi = vld1_s16(kSinTab1);
kSinTab1 += 4;
int16x4_t tmpr = vld1_s16(kSinTab2);
kSinTab2 -= 4;
int16x4_t inre_0 = vld1_s16(inre1);
inre1 += 4;
int16x4_t inre_1 = vld1_s16(inre2);
inre2 -= 4;
int16x4_t inim_0 = vld1_s16(inim1);
inim1 += 4;
int16x4_t inim_1 = vld1_s16(inim2);
inim2 -= 4;
tmpr = vneg_s16(tmpr);
inre_1 = vrev64_s16(inre_1);
inim_1 = vrev64_s16(inim_1);
tmpr = vrev64_s16(tmpr);
int32x4_t xr = vmull_s16(tmpr, inre_0);
int32x4_t xi = vmull_s16(tmpr, inim_0);
int32x4_t yr = vmull_s16(tmpr, inim_1);
int32x4_t yi = vmull_s16(tmpi, inim_1);
xr = vmlal_s16(xr, tmpi, inim_0);
xi = vmlsl_s16(xi, tmpi, inre_0);
yr = vmlal_s16(yr, tmpi, inre_1);
yi = vmlsl_s16(yi, tmpr, inre_1);
yr = vnegq_s32(yr);
xr = vshrq_n_s32(xr, 5);
xi = vshrq_n_s32(xi, 5);
yr = vshrq_n_s32(yr, 5);
yi = vshrq_n_s32(yi, 5);
int32x4_t outr0 = vsubq_s32(xr, yi);
int32x4_t outr1 = vaddq_s32(xr, yi);
int32x4_t outi0 = vaddq_s32(xi, yr);
int32x4_t outi1 = vsubq_s32(yr, xi);
// Find the absolute maximum in the vectors.
int32x4_t tmp0 = vabsq_s32(outr0);
int32x4_t tmp1 = vabsq_s32(outr1);
int32x4_t tmp2 = vabsq_s32(outi0);
int32x4_t tmp3 = vabsq_s32(outi1);
// vabs doesn't change the value of 0x80000000.
// Use u32 so we don't lose the value 0x80000000.
max_r = vmaxq_u32(max_r, vreinterpretq_u32_s32(tmp0));
max_i = vmaxq_u32(max_i, vreinterpretq_u32_s32(tmp2));
max_r = vmaxq_u32(max_r, vreinterpretq_u32_s32(tmp1));
max_i = vmaxq_u32(max_i, vreinterpretq_u32_s32(tmp3));
// Store the vectors.
outr1 = vrev64q_s32(outr1);
outi1 = vrev64q_s32(outi1);
int32x4_t outr_1 = vcombine_s32(vget_high_s32(outr1), vget_low_s32(outr1));
int32x4_t outi_1 = vcombine_s32(vget_high_s32(outi1), vget_low_s32(outi1));
vst1q_s32(outre1, outr0);
outre1 += 4;
vst1q_s32(outim1, outi0);
outim1 += 4;
vst1q_s32(outre2, outr_1);
outre2 -= 4;
vst1q_s32(outim2, outi_1);
outim2 -= 4;
}
max_r = vmaxq_u32(max_r, max_i);
#if defined(WEBRTC_ARCH_ARM64)
uint32_t maximum = vmaxvq_u32(max_r);
#else
uint32x2_t max32x2_r = vmax_u32(vget_low_u32(max_r), vget_high_u32(max_r));
max32x2_r = vpmax_u32(max32x2_r, max32x2_r);
uint32_t maximum = vget_lane_u32(max32x2_r, 0);
#endif
return (int32_t)maximum;
}
